Temporal-aware semi-supervised performance prediction and resource allocation for cloud-native microservices

Cloud-native loosely coupled microservices support elastic and scalable deployments while accelerating development cycles and reducing update complexity. However, due to scarce labeled performance data and complex inter-node interactions, optimizing resource allocation remain significant challenges. Accordingly, we propose TSPRA (Time-aware Semi-supervised Performance Prediction and Resource Allocation), a novel resource adjustment method for microservice systems. Specifically, TSPRA first collects limited labeled resource-performance data and augments the dataset using TabDDPM, a denoising diffusion probabilistic model tailored for tabular data. TSPRA then constructs a temporal-aware semi-supervised prediction model (TSMD) that captures resource usage patterns through contrastive learning and effectively predicts microservice response times. When predictions indicate that current resource configurations may violate QoS targets, a hyper-heuristic resource allocation model (HHA) is activated to adaptively adjust CPU, memory, and I/O allocations. HHA employs a two-tier optimization framework: the high-level Ant Colony Optimization (ACO) dynamically schedules three complementary low-level heuristics—Whale Optimization Algorithm (WOA), Quantum-behaved Particle Swarm Optimization (QPSO), and Simulated Annealing (SA)—to search for optimal resource configurations that minimize end-to-end latency while maximizing resource utilization. The simulation tests based on Docker and DeathStarBench verify the effectiveness of our methods.